问题
Fast Fourier Transforms (FFTs) in Numpy/SciPy are not threaded. Enthought Python is shipped with the Intel MKL numerical library, which is capable of threaded FFTs. How does one get access to these routines?
回答1:
The following code works for me with Enthought 7.3-1 (64-bit) on Windows 7 Ultimate 64-bit. I haven't benchmarked it but it certainly uses all cores at once rather than just one.
from ctypes import *
class Mkl_Fft:
c_double_p = POINTER(c_double)
def __init__(self,num_threads=8):
self.dfti = cdll.LoadLibrary("mk2_rt.dll")
self.dfti.MKL_Set_Num_Threads(num_threads)
self.Create = self.dfti.DftiCreateDescriptor_d_md
self.Commit = self.dfti.DftiCommitDescriptor
self.ComputeForward = self.dfti.DftiComputeForward
def fft(self,a):
Desc_Handle = c_void_p(0)
dims = (c_int*2)(*a.shape)
DFTI_COMPLEX = c_int(32)
rank = 2
self.Create(byref(Desc_Handle), DFTI_COMPLEX, rank, dims )
self.Commit(Desc_Handle)
self.ComputeForward(Desc_Handle, a.ctypes.data_as(self.c_double_p) )
Usage:
import numpy as np
a = np.ones( (32,32), dtype = complex128 )
fft = Mkl_Fft()
fft.fft(a)
回答2:
A cleaner version of my original answer is as follows:
from ctypes import *
mkl = cdll.LoadLibrary("mk2_rt.dll")
c_double_p = POINTER(c_double)
DFTI_COMPLEX = c_int(32)
DFTI_DOUBLE = c_int(36)
def fft2(a):
Desc_Handle = c_void_p(0)
dims = (c_int*2)(*a.shape)
mkl.DftiCreateDescriptor(byref(Desc_Handle), DFTI_DOUBLE, DFTI_COMPLEX, 2, dims )
mkl.DftiCommitDescriptor(Desc_Handle)
mkl.DftiComputeForward(Desc_Handle, a.ctypes.data_as(c_void_p) )
mkl.DftiFreeDescriptor(byref(Desc_Handle))
return a
def ifft2(a):
Desc_Handle = c_void_p(0)
dims = (c_int*2)(*a.shape)
mkl.DftiCreateDescriptor(byref(Desc_Handle), DFTI_DOUBLE, DFTI_COMPLEX, 2, dims )
mkl.DftiCommitDescriptor(Desc_Handle)
mkl.DftiComputeBackward(Desc_Handle, a.ctypes.data_as(c_void_p) )
mkl.DftiFreeDescriptor(byref(Desc_Handle))
return a
回答3:
New and improved version which handles arbitrary strides in the input and output arrays. By default this one is now not-in-place and creates a new array. It mimics the Numpy FFT routines except that it has a different normalisation.
''' Wrapper to MKL FFT routines '''
import numpy as _np
import ctypes as _ctypes
mkl = _ctypes.cdll.LoadLibrary("mk2_rt.dll")
_DFTI_COMPLEX = _ctypes.c_int(32)
_DFTI_DOUBLE = _ctypes.c_int(36)
_DFTI_PLACEMENT = _ctypes.c_int(11)
_DFTI_NOT_INPLACE = _ctypes.c_int(44)
_DFTI_INPUT_STRIDES = _ctypes.c_int(12)
_DFTI_OUTPUT_STRIDES = _ctypes.c_int(13)
def fft2(a, out=None):
'''
Forward two-dimensional double-precision complex-complex FFT.
Uses the Intel MKL libraries distributed with Enthought Python.
Normalisation is different from Numpy!
By default, allocates new memory like 'a' for output data.
Returns the array containing output data.
'''
assert a.dtype == _np.complex128
assert len(a.shape) == 2
inplace = False
if out is a:
inplace = True
elif out is not None:
assert out.dtype == _np.complex128
assert a.shape == out.shape
assert not _np.may_share_memory(a, out)
else:
out = _np.empty_like(a)
Desc_Handle = _ctypes.c_void_p(0)
dims = (_ctypes.c_int*2)(*a.shape)
mkl.DftiCreateDescriptor(_ctypes.byref(Desc_Handle), _DFTI_DOUBLE, _DFTI_COMPLEX, _ctypes.c_int(2), dims )
#Set input strides if necessary
if not a.flags['C_CONTIGUOUS']:
in_strides = (_ctypes.c_int*3)(0, a.strides[0]/16, a.strides[1]/16)
mkl.DftiSetValue(Desc_Handle, _DFTI_INPUT_STRIDES, _ctypes.byref(in_strides))
if inplace:
#Inplace FFT
mkl.DftiCommitDescriptor(Desc_Handle)
mkl.DftiComputeForward(Desc_Handle, a.ctypes.data_as(_ctypes.c_void_p) )
else:
#Not-inplace FFT
mkl.DftiSetValue(Desc_Handle, _DFTI_PLACEMENT, _DFTI_NOT_INPLACE)
#Set output strides if necessary
if not out.flags['C_CONTIGUOUS']:
out_strides = (_ctypes.c_int*3)(0, out.strides[0]/16, out.strides[1]/16)
mkl.DftiSetValue(Desc_Handle, _DFTI_OUTPUT_STRIDES, _ctypes.byref(out_strides))
mkl.DftiCommitDescriptor(Desc_Handle)
mkl.DftiComputeForward(Desc_Handle, a.ctypes.data_as(_ctypes.c_void_p), out.ctypes.data_as(_ctypes.c_void_p) )
mkl.DftiFreeDescriptor(_ctypes.byref(Desc_Handle))
return out
def ifft2(a, out=None):
'''
Backward two-dimensional double-precision complex-complex FFT.
Uses the Intel MKL libraries distributed with Enthought Python.
Normalisation is different from Numpy!
By default, allocates new memory like 'a' for output data.
Returns the array containing output data.
'''
assert a.dtype == _np.complex128
assert len(a.shape) == 2
inplace = False
if out is a:
inplace = True
elif out is not None:
assert out.dtype == _np.complex128
assert a.shape == out.shape
assert not _np.may_share_memory(a, out)
else:
out = _np.empty_like(a)
Desc_Handle = _ctypes.c_void_p(0)
dims = (_ctypes.c_int*2)(*a.shape)
mkl.DftiCreateDescriptor(_ctypes.byref(Desc_Handle), _DFTI_DOUBLE, _DFTI_COMPLEX, _ctypes.c_int(2), dims )
#Set input strides if necessary
if not a.flags['C_CONTIGUOUS']:
in_strides = (_ctypes.c_int*3)(0, a.strides[0]/16, a.strides[1]/16)
mkl.DftiSetValue(Desc_Handle, _DFTI_INPUT_STRIDES, _ctypes.byref(in_strides))
if inplace:
#Inplace FFT
mkl.DftiCommitDescriptor(Desc_Handle)
mkl.DftiComputeBackward(Desc_Handle, a.ctypes.data_as(_ctypes.c_void_p) )
else:
#Not-inplace FFT
mkl.DftiSetValue(Desc_Handle, _DFTI_PLACEMENT, _DFTI_NOT_INPLACE)
#Set output strides if necessary
if not out.flags['C_CONTIGUOUS']:
out_strides = (_ctypes.c_int*3)(0, out.strides[0]/16, out.strides[1]/16)
mkl.DftiSetValue(Desc_Handle, _DFTI_OUTPUT_STRIDES, _ctypes.byref(out_strides))
mkl.DftiCommitDescriptor(Desc_Handle)
mkl.DftiComputeBackward(Desc_Handle, a.ctypes.data_as(_ctypes.c_void_p), out.ctypes.data_as(_ctypes.c_void_p) )
mkl.DftiFreeDescriptor(_ctypes.byref(Desc_Handle))
return out
来源：https://stackoverflow.com/questions/11752898/threaded-fft-in-enthought-python